Optical vs Electrical Tradeoffs for Communications in Highly Parallel, Distributed Memory Multiprocessors

Abstract

Continued scaling of silicon CMOS technologies to smaller feature sizes will provide impressive opportunities for integration of very compex processing systems on single ICs and single wafers. This suggests a considerable scaling of "large-scale" systems such as distributed computing environments to much smaller, more highly monolithic real-izations. Such "scaled systems" will require a corresponding scaling of communications networks, perhaps reduced to single wafer sized networks and switching circuits. Presently, electronic interconnections characterize monolithic circuit interconnections while optical communications is making significant inroads for efficient realization of system communications. In the future "scaled" world of ULSI systems, it remains unclear whether optical interconnection techniques increasingly favored for large-scale system communications will remain favored. However, the system organizations and operating protocols are likely to change significantly as present large-scale systems are scaled. It is suggested here that a principle objective will be to develop networks with disposable bandwidth, well in excess of the actual bandwidth required. Under such conditions several of the control issues limiting efficient use of distributed systems are greatly relaxed. Optical interconnects, integrated for example on silicon wafer substrates, provide the high bandwidth sought. However, resynchronization delays and queues will be very difficult in the optical domain. It is suggested here that the overall network function can be divided into relatively slow electronic control of the network paths and high bandwidth (optical or electrical) links and switches, providing a high performance, hybrid network realization.